find_events() returns times with offset

I have a FIF file with a STI101 trigger channel to which I apply mne.find_events() to extract the event markers. It seems to be that there is a mismatch of 5100 samples between the extracted events and the positions of the events when looking at the STI-Channel manually. Am I missing something? Why is there a delay to the raw STIM channel positions?

sample file:

import numpy as np
import mne

raw = mne.io.read_raw('events_raw.fif', preload=True)
# automatic extraction
events = mne.find_events(raw,stim_channel='STI101', shortest_event=1, consecutive=False)
# "manual" exctraction of events
manual_events = np.nonzero(raw.get_data('STI101'))[-1]
manual_events = manual_events[np.diff(np.pad(manual_events, [1, 0]))>1]

# consistent lag of 5100 samples
print(events[:,0]-manual_events)

however, I cannot reproduce this when I create the the same structure with an RawArray, which leads me to conclude that this is something inside of my file that I just can’t figure out, but would badly need to fix.

raw = mne.io.read_raw('events_raw.fif', preload=True)
data = raw.get_data()
ch_types = ['grad', 'stim', 'syst']
ch_names = ['MEG2633', 'STI101', 'SYS201']
info = mne.create_info(ch_names=ch_names, sfreq=1000.0, ch_types=ch_types)
raw = mne.io.RawArray(data, info)
events = mne.find_events(raw,stim_channel='STI101', shortest_event=1, consecutive=False)

# now it works perfectly fine as expected, 0 lag
print(events[:,0]-manual_events)

Do you have any idea why this is happening?

Additional information

Platform:      Windows-10-10.0.19041-SP0
Python:        3.8.8 (default, Feb 24 2021, 15:54:32) [MSC v.1928 64 bit (AMD64)]
Executable:    C:\Users\skjerns\anaconda3\python.exe
CPU:           AMD64 Family 23 Model 113 Stepping 0, AuthenticAMD: 24 cores
Memory:        31.9 GB

mne:           0.22.1
numpy:         1.19.2 {blas=mkl_rt, lapack=mkl_rt}
scipy:         1.6.2
matplotlib:    3.3.4 {backend=Qt5Agg}

sklearn:       0.24.1
numba:         0.53.1
nibabel:       Not found
nilearn:       Not found
dipy:          Not found
cupy:          Not found
pandas:        1.2.3
mayavi:        Not found
pyvista:       Not found
vtk:           Not found

[agramfort]

https://mne.tools/stable/glossary.html?highlight=first_samp

[skjerns]

Is there a place where I can find more documentation on first_samp, and how it is used, or what its definition is?

I’m a bit puzzled because I have the case in multiple files that events[-1,0] > len(raw). In my understanding the fif contains all samples, and first_samp tells mne at which position the actual data starts, right? So how can my events be out of bounds for raw.get_data()? Or is it possible that my fif files are already cropped (by e.g. MaxFilter or ELECTRA), and that’s why first_samp has been noted in the file, and raw.get_data() returns an already cropped array? So first_samp is only used to annotate how many samples there were before the current file (which would be a bit odd to me)?